# This is an example that demonstrates how to configure a model file.
# You can modify the configuration according to your own requirements.

# to print the register_table:
# from funasr.register import tables
# tables.print()

# network architecture
model: LLMASR2
model_conf:
    lsm_weight: 0.1     # label smoothing option
    length_normalized_loss: true

# encoder
audio_encoder: "/nfs/zhifu.gzf/init_model/SenseVoiceModelscope"
audio_encoder_conf:
    hub: ms
    freeze: true

llm: Qwen1.5-7b-chat
llm_conf:
  hub: hf
  freeze: true
  init_param_path: "/nfs/zhifu.gzf/init_model/qwen/Qwen1___5-7B-Chat_raw"

audio_adaptor: Transformer
audio_adaptor_conf:
  downsample_rate: 2
  llm_dim: 4096
  encoder_dim: 1280
  n_layer: 2

# frontend related
frontend: WhisperFrontend
frontend_conf:
    fs: 16000
    whisper_model: large-v3
    do_pad_trim: false
    permute: false # true: [bs, frames, dims]; false: [bs, dims, frames]
    filters_path: "/nfs/zhifu.gzf/init_model/SenseVoiceModelscope/assets/mel_filters.npz"



train_conf:
  accum_grad: 1
  grad_clip: 5
  max_epoch: 15
  keep_nbest_models: 10
  log_interval: 10

optim: adamw
optim_conf:
   lr: 0.0001
   weight_decay: 0.000000

scheduler: warmuplr
scheduler_conf:
   warmup_steps: 1500

dataset: OpenAIDataset
dataset_conf:
    index_ds: OpenAIIndexDSJsonl
    batch_sampler: CustomDistributedBatchSampler
    batch_type: example # example or length
    batch_size: 4 # if batch_type is example, batch_size is the numbers of samples; if length, batch_size is source_token_len+target_token_len;
    max_token_length: 3000 # filter samples if source_token_len+target_token_len > max_token_length,
    shuffle: True
    num_workers: 0
    audio_adaptor_downsample_rate: ${audio_adaptor_conf.downsample_rate}
    audio_encoder_downsample_rate: 2
#    prompt: "<|startoftranscription|><|zh|><|transcribe|><|zh|><|notimestamps|><|wo_itn|>"



tokenizer: HuggingfaceTokenizer
tokenizer_conf:
  init_param_path: "/nfs/zhifu.gzf/init_model/qwen/Qwen1___5-7B-Chat_raw"